Real-time Object Detection from Ground Penetrating Radar Images Using YOLOv7

Abstract

In this paper, we examine the presence of subsurface objects and the detection of four types of objects from ground penetrating radar (GPR) images obtained by a webcam using YOLOv7 in real-time. In this study, four types of objects (styrofoam, wood, concrete, and aluminum) were buried in a sandbox, and 64 radar images were generated using 2600 MHz radar for learning and detection by YOLOv7. The 64 images were augmented to 192 using the Cutout method, of which 168 were used for training and 24 for validation. In addition to the 192 images for training and verification, 64 test images were generated and cross-validated ten times to evaluate real-time detection by the webcam. The result is an accuracy of about 89% for the presence of objects and a detection rate of 39-83% for four types of objects.